This is a mixture-of-experts inference engine written in pure C, with zero dependencies and no GPU required. It keeps only the parts the model needs on every run in memory, and reads the remaining expert weights from disk on demand. With as little as 8GB of RAM, it can run open-source large models such as GLM, Kimi, DeepSeek, and Qwen, and provides OpenAI- and Anthropic-compatible APIs. It works on Windows, Linux, and macOS.